AI 生成广告涌入中国信息流，平台以 AI 审核反击

（本文作者为 Chelsea_Sun，钛媒体经授权发布）

NextFin News — A skincare pitch built from a cloned face and voice is the kind of abuse users notice. Most of the mess is quieter. Generative tools have made short ad videos almost free. A fifteen-second product clip can cost little more than lunch money. A live shoot—even a budget one with a day-rate model, camera, and lights—used to run far higher per cut. Shops that never bought creative can buy it now. So can outfits that care more about volume than about rules.

That change breaks the old review playbook. Ads used to be written by people, checked by people, and judged against rules people could read. Machines now draft the inventory. Gray workshops can stamp out fake endorsements and borrowed authority at scale. The law still leans on case-by-case human logic. The feed does not wait.

Some bad ads look ordinary still by still. Platform safety teams point to short animated spots built around social flashpoints—bridal gift fights and similar bait—that pass a single-image check and fail only when picture, caption, and sound are read together. Older detectors that score frames miss the story. Reviewers need models that watch the whole clip.

Douyin’s ads unit, Ocean Engine, has pushed hard on that shift, as has Xiaohongshu. Review is moving toward agent-style stacks: one main system for broad risk, smaller ones for likeness theft, digital humans, fakes, and the like. The company says interception capacity has jumped by an order of magnitude. Officials there talk about making AI the default review layer within about a year, with rules refreshed monthly. The volume leaves little choice. Industry research they cite finds most ad buyers already use AI for some or all production, and a large share of creatives are synthetic.

Generation and policing now race on the same track. What got cheaper to make also got cheaper to abuse. Catching it takes more compute, better models, and tighter pipelines—not another PDF of guidelines.

Tighter filters create their own tax. Agencies say each platform runs different rules, staff burn hours tracking changes, and rejections still land without a clear reason. How fast an appeal is answered, they argue, is part of whether a platform is fit to take money. Ocean Engine staff who meet advertisers hear a sharper complaint than any one refusal: people do not know why a cut failed, how to fix it, how to appeal, or how to stay clean next time.

So the same generative toolkit is being turned on the help desk. Rejection notes try to point to the problem, name the rule, and suggest a fix. One-click repair is meant to turn that note into an edit. Rule pages are supposed to spell out the basis, the date, and the required change in language a buyer can learn. None of that wipes out false blocks. It can cut the cost of being honest.

The law does not fix a number on how much computing power a platform must spend on clean ads. The market will. If every sponsored item forces the viewer to double-check the face, the claim, and the brand, advertisers are buying weaker attention. That is not only a compliance issue. It is maintenance of a business in which ads still work.

Synthetic creative is not going away. The live question is whether detection can keep up—and whether platforms treat the race as core plumbing or as a cost to squeeze. Advertisers already feel the answer in their rejection queues. Users will give theirs more slowly: by whether they still believe the feed.

更多精彩内容，关注钛媒体微信号（ID：taimeiti），或者下载钛媒体App